Preserve missing-vote candidates in final-stack survivors

[seathatflowsinourveins]

Scope

Preserve unresolved candidates when a sweep's only refutation is a missing return, using the existing native ledger classifier. Cache parsed vote documents once per build, and treat a later covering sweep's missing-vote-only re-proposal as reopening an earlier refutation, as ruled for Chandra.

• Head: 5f6c068a9650c0df0c907483f5536931fbcb10ca.
• Main pin: 8b1ef0c37e2f523aa84d2c75960540f45e74e814; lane: lane:foundation.
• Reviewed 1916274b was replayed as 50e10f77fed7b6107ea6f622af904e58760e2f14; ONE correction follows that rebased tip.
• Owned paths: generator, existing test module, generated final-stack report and their three evidence bindings.

SOTA sources

• Native ledger at main 8b1ef0c3: refutes_on_merit, refuted_by_absence, ref_resolver at 657–658, and Checker.document at 733–736. Follow its existing per-path parsed-document cache pattern, retaining safe loading and first-read source hashes. The classifier/resolver are unchanged.
• Protocol step 2 at that main: cumulative unresolved survivors persist; missing returns do not settle them. The CC's 2026-10-12 ruling applies this to a later missing-vote-only re-proposal after a prior refutation.
• Ledger contract: retained missing markers identify refutations by absence. Existing unittest fixtures test public generator output and observe the file-read boundary; no new runner or classifier is introduced.

Evidence-class table

Claim	Evidence class	Returned evidence
Repeated vote reads fail before caching	synthetic, local_integration	Reviewed generator bytes: one test/one failure/zero errors, rc 1, 0.006s; four reads versus one expected
Missing-only Chandra re-proposal fails before correction	synthetic, local_integration	Reviewed generator bytes: one test/one failure/zero errors, rc 1, 0.007s; missing survivor
Both regressions pass after correction	local_integration	Two named tests, rc 0, 0.013s; full generator module 75 tests/158.447s/zero skips, rc 0
Cache removes repeated parsing	local_integration	Native build 9.150725s before and 0.122853s with cache only; vote-document reads 245/162/9 → 1/1/1; cache-only output equals the rebased report. Single observations under host load, not a benchmark distribution
Integrity and native consumers pass	local_integration	FULL validate rc 0: 70 components/11,738 files/four profiles/239 receipts; ledger module 99 tests/20.560s and validate module 86 tests/4.656s, both rc 0; generator/evidence checks rc 0
Publication and bounded scans pass	native_proven	Native report writer and changed-file scan rc 0; four files/zero host findings; gitleaks two-commit PR range rc 0/zero findings

Local commands run

python3 -B -m unittest tests.test_final_stack.StatusTests.test_shared_vote_document_is_read_once_per_build
python3 -B -m unittest tests.test_final_stack.StatusTests.test_reproposed_chandra_with_only_a_missing_vote_reopens_a_prior_refutation
# Each fails against unchanged reviewed generator bytes before its fix, then both pass.
python3 -B scripts/final_stack.py --write
python3 -B scripts/final_stack.py --check
python3 -B scripts/validate.py
python3 -B -m unittest tests.test_final_stack
python3 -B -m unittest tests.test_saturation_ledger
python3 -B -m unittest tests.test_validate
gitleaks git --redact=100 --config .gitleaks.toml --log-opts=8b1ef0c37e2f523aa84d2c75960540f45e74e814..HEAD
# All return 0. Test counts overlap named fixtures and are not summed.

Generated fields and rebase

The original correction restores 21 survivors. The CC's Chandra ruling adds one more: document-retrieval 12 → 13, overall 106 → 107, dated at its 2026-09-29 re-proposal. A prior survivor such as Exa retains its existing provenance. A missing entry without a later re-proposal does not reopen a refutation.

Against the rebased reviewed tip, changes are the Chandra survivor, document-retrieval surviving/refuted funnel counts, its repository index role, summary count and the corresponding definitions. Runs, winners, statuses and source bindings remain unchanged by the correction. The rebase carries main's updated skills-lifecycle source hash. The report is 309,186 bytes, SHA256 5dde8ad61bb124fbcb2d56acb6c0973febc0802fda54bf5e5d8127ce9c15153b.

Rebase range-diff shows one = outside the evidence registry. Native stable patch IDs from zero-context diffs also match outside that registry; the full range-diff differs only in the recomputed report binding. Main's 11,738 evidence entries and non-file records are retained; only the three owned bindings change and match actual bytes.

Decision record and thread disposition

The existing protocol and the CC's 2026-10-12 adjudication govern this correction. Codex thread PRRT_kwDOUg_LrM6rTSKs identifies the reproduced caching defect; it has been replied to with the fix and resolved. The correction uses a build-scoped parsed-document cache and keeps the existing vote resolver/classifier. The regression verifies one read per document per build, including a second build with an independent cache and matching source hash. Chandra's separate fail-first fixture covers a prior substantive lens refutation followed by a missing-only re-proposal.

Follow-ups

• Astra delta P3-1 / CC I2: represent the retained 2026-10-03 GPT fit supplement's 27 decisive entries through the agreed supplement ledger record type or append them as records, then regenerate and test ledger/report bindings; candidate selections remain separate from this fix.
• Claude P3: add prior-survival/later-missing-return fixtures preserving the original sweep/date, including a verdict between those sweeps; the real Exa history remains the current provenance control.
• Astra delta P3-2: add an unavailable-retained-vote fixture and propagate a named unresolved-reference recovery error before publication, retaining the shared classifier for resolvable references; this is the inherited recovery boundary.
• Astra delta P3-3: clean historical machine-name hygiene separately, updating affected artifact paths and bindings together; this patch adds no hostname occurrences.
• Astra delta P3-4: preserve Git executable modes in merged review overlays or supply a mode manifest; the delta packet omitted one live-base-added executable mode, corrected by the reviewer. This is packet construction work, not a PR source defect.
• Astra delta P3-5: attach identical copies and hashes for the three retained returns and three cited first-party source files in review extra/, or explicitly designate their authenticated tree/Git blobs as retained source copies; published bytes were verified, while original capture provenance remains unverified.
• Astra delta P3-6: cache-source naming corrected in this body from Validator.document to the pinned Checker.document; the cited per-path cache implementation and code behavior are unchanged.
• Astra delta P3-7: the survived versus missing-vote-pending labelling for observation-inference (6), recovery-portability (1) and token-efficiency (5) landed with docs: close seven W6 foundation layers with preregistered comparisons #1023 (a365d8b). What remains is including all six ci-supply-chain survivors in that layer's W6 closeout, which stays with its owner.
• Claude micro P3-1 at 5f6c068: two of the new guard's three conditions cannot be false on a ledger that passes its own checker; pin the earlier-sweep comparison with a test, or simplify the guard.

Host evidence and checklist

No host receipts or platform statuses change. Fixture and build measurements establish repository integration, not provider/GPU/adoption or vendor-harness acceptance. Astra ACK and Claude CHANGES_REQUESTED at 1916274 are recorded; their delta confirmations at 5f6c068 are Astra ACK and Claude micro PASS_WITH_P3.

ONE exact-1916274b leased push returns 0. Native pre-push scans 12,537 files with zero host matches; three registry tests pass in 0.267s and one published-digest privacy test passes in 12.951s. Main advanced after the rebase; read-only merge-tree against observed 421bf3f98b631d931838c22a927b4fec246f7df8 is clean, but local acceptance is pinned to the rebased main above.

• Rebase precedes ONE correction commit; both new regressions fail first.
• The native cache pattern, and the classifier, resolver and safe source loading, are reused.
• Regenerated report, FULL validate, owning modules and bounded scans pass.
• No credentials, client configuration, protected paper state or peer worktrees changed.
• Reviewer P3s and unrecorded supplement remain follow-ups.
• Exact-head hosted checks pass at 5f6c068; the Astra delta returned ACK and the Claude micro PASS_WITH_P3. Landing is the command center's step.

Follow-ups after landing

Landed at 2f79ae7 on 2026-10-12T13:51:33Z from head 5f6c068 (Claude micro PASS_WITH_P3, GPT read ACK; Codex unavailable at its usage limit). The Follow-ups listed above carry forward unchanged; none blocks.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-12T00:59:33.802971Z	1916274	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1916274b77

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[seathatflowsinourveins]

@codex review

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.